Simple references explaining problems of exploration of unknown space


#1

I am now working in a chemistry lab trying to use principle of developmental robotics and machine learning in general for the exploration of chemical spaces.

I have been trying to explain why random or grid search experiments are not the best things to do for some time now, but it does not seem to be fully understood yet. I have been explaining the problem by referring to injective and surjective functions and drawing a lot of example but we are not yet there :slight_smile:

Do you know of any good website or papers with visualisation that explain the problem of exploration of an unknown function/system?

I know this is all natural to us and we always refer to it as obvious in papers but somehow I could no find a simple visually made paper to explain this to non computer scientists.

More concretely is there any paper from the group explaining the figure below in simple terms?


(Clement Moulin Frier) #2

Hi @jgrizou

I actually tried to explain these points in my ICDL2013 paper and its associated presentation.

In the presentation, there is an attempt at representing the problem visually on Slide 2 and Slide 4 (slide numbers indicated in the page footers, not the pdf page numbers).

Slide 4 illustrates how goal babbling allows to obtain a more uniform distribution on the output space compared to motor babbling. The two last animations of this slide are are illustrating that usually the bounds of the output space are not known, and that parts of the function can be unlearnable (or less learnable, e.g. noisy). These two last points favor the use of active exploration strategies instead of random ones (e.g., random goal babbling performs badly if the output space bounds are not known or if the “function” is noisy, hence the interest of active approaches like learning progress maximization).

Don’t hesitate to reuse this material if you want, or ask more detail if needed.

Cheers,
Clément


(Pierre-Yves Oudeyer) #3

Hello,

I think another good place to find nice visualizations for a tutorial on exploration/goal babbling is chapter 0 of @benureau 's PhD thesis: http://fabien.benureau.com/docs/phd_benureau.pdf

Pierre-Yves